Epoch AI’s Capabilities Index shows both model families making substantial gains since 2024: GPT rose 40.2 points from its best 2024 GPT-4o snapshot to GPT-6 Astra, while Claude climbed 42.8 points from Claude 3 Opus to Opus 5.5.

Using the first 2024 flagship for each lab, the article calculates annualized gains of about 17.7 index points for GPT and 16.7 for Claude; those rates, like the overall gains, are the authors’ arithmetic.

GPT-6 Astra scored 169.1, just 0.7 points ahead of Opus 5.5 at 168.4, a slim margin that does not settle the comparison.

The models’ 90% score intervals overlap—166.16–172.28 for Astra and 165.54–171.51 for Opus 5.5—and Epoch AI previously assigned a 72% probability that Astra would rank higher.

Across seven tests scored for both newest models, Astra leads on five, including graduate-level science, factual recall, chess puzzles and ARC-AGI-2, and ties Opus 5.5 on competition math and ARC-AGI-1.

Looking at four shared tests from each lab’s 2024 flagship to its newest model, Claude made larger gains in factual recall and competition math, while GPT improved more on chess puzzles and graduate-level science; including the index, Claude gained more in three of five comparisons and GPT in two.

The comparison has important limits: the labs do not share every benchmark, some tests have ceiling scores, and each result reflects the model’s best tested reasoning setting rather than performance in a particular product or at a given price; a separate AI index ranks Opus 5.5 first.

The figures come from Epoch AI’s public benchmarking data, downloaded Nov. 8, 2026, and compiled by Alex Michael; the article says it checked the cited results against Epoch’s data, which are not self-reported by the labs.